表紙画像生成

抄録

The purpose of this research is to make a system to design book covers easily. For this purpose, we trained a deep learning-based model to generate the images from a layout graph, where book cover elements, such as the title text information and some visual objects, correspond to its nodes and the geometric relationships among elements specify its edges. To make the generated book cover images more realistic, we used a compositive loss function involving the quality of each object and the difference from the ground truth images.